Application And Research Of Fault Detection Based On Convolutional Neural Network

During seismic exploration,seismic data interpretation is a critical part of structural description,reservoir prediction and favorable exploration,the quality of the interpretation directly affects the development of oil and gas exploration,while the fault interpretation is the core and foundation of seismic data interpretation.During the past,to identify faults is usually based on the different manifestations of faults in seismic data.However,such an interpretation method has the disadvantages of cumbersome processes,long time-consuming,large errors,and reliance on personal interpretation experience.Also lacks certain objectivity.Therefore,how to identify faults efficiently and accurately has become a problem that researchers are very concerned about.As there are so many problems in artificial fault interpretation,experts and scholars,domestic and overseas have proposed much methods to extract fault attribute,such as coherence attribute,variance attribute,ant body attribute,curvature attribute,etc.they all show the fault information that hidden in the seismic data,thus could reduce the randomness of the interpretation results.Recent years,the neural network makes a great progress in the field of image processing.And the structure of seismic data is similar to the image.Therefore,the successful application of the neural network on the image processing can be used to identify faults,which makes it a new research direction.In this paper,a traditional convolutional neural network is constructed.Then,two methods of obtaining training data are explained: one is manual interpretation of seismic data,the other is to make synthetic seismic data that containing different types of faults.In order to further increase the diversity of training data to improve the generalization of the model,data enhancement is carried out.Affected by the magnitude of seismic data,gradients are difficult to transmit effectively during model training,and it is difficult for the network to converge to the optimal solution.Therefore,this paper normalized the seismic data to determined range.Some model parameters were discussed during the model training process,and the optimal model parameters were determined.Then some synthetic seismic data are used to test the model,it is verified that convolutional neural networks are feasible to identify faults in seismic data.However,traditional convolutional networks have the disadvantages of repeated calculations,low efficiency,and low precision,so this paper proposes a U-NET-based end-to-end full convolutional network to identify faults.After test in realistic seismic data,it shows that the U-NET-based network has better performance compared with traditional convolutional network whether the computational efficiency or accuracy,achieve more efficient and accurate fault identification.